High-flow nasal oxygen therapy compared with conventional oxygen therapy in hospitalised patients with respiratory illness: a systematic review and meta-analysis

Search strategy and study selection

A comprehensive search was performed on 29 June 2021, and subsequently updated on 22 March 2023, for articles within the last 50 years across Embase, Medline via Ovid and CENTRAL databases. The full search strategy and keywords employed can be found in online supplemental table 1A–F. The study protocol was registered with PROSPERO (CRD42021264837).

To meet inclusion criteria, studies had to evaluate HFNO in hospitalised adult patients (≥18 years) with acute respiratory illness necessitating oxygen therapy, indicating ARF. RCTs, cross-over studies and observational studies with an intervention and comparator group were included. Hospitalisation was defined as ED presentation or admission to an ICU or ward. HFNO was defined as flow rates >20 L/min delivered via specialised high-flow devices and nasal cannulae. COT was defined as flow rates ≤15 L/min delivered via standard nasal cannulae, simple face mask, non-rebreather reservoir mask or venturi mask. Illnesses causing ARF (as defined by study authors and with or without hypercapnia) included upper and lower respiratory tract infections, COVID-19, acute respiratory distress syndrome, acute exacerbations of a chronic respiratory illness, pulmonary embolism, pneumothorax, acute lung injury and transfusion-related acute lung injury. Studies examining patients with stable, chronic respiratory illnesses (without a concurrent acute exacerbation), with patients under 18 years old, comparing HFNO to non-invasive ventilation (NIV) or invasive mechanical ventilation (IMV) only or written in languages other than English were excluded.

The primary outcome for this review was hospital mortality rate, with secondary outcomes including short-term mortality (≤30 days), long-term mortality (>30 days), escalation to IMV, escalation to NIV, ICU and hospital length of stay (LOS), changes in partial pressure of carbon dioxide (PaCO₂) levels, disability as defined by study authors (eg, dyspnoea scores) and inpatient admission rates for studies evaluating patients in ED.

Data extraction

Two independent reviewers (DS, YHK, S-WK, AP, DMS, YN or NS) screened abstracts and full-text articles sequentially in pairs for eligible studies. A third reviewer (YHK or NS) resolved disagreements between reviewer pairs by adjudication. For each included study, the risk of bias was assessed independently by two authors (DS, AP or S-WK) with adjudication by a third author (YHK or NS). RCTs and randomised cross-over studies were assessed using the revised Cochrane risk-of-bias tool for randomised trials (RoB 2) while non-randomised interventional studies were assessed using the risk of bias in non-randomised studies of interventions (ROBINS-I) framework. The Newcastle-Ottawa Scale was used for assessments of observational studies.

Data synthesis

Treatment effects were pooled using the random effects model to weigh heterogeneous studies more equally than the fixed effects model, as many studies were clinically and methodologically heterogeneous. The Mantel-Haenszel method was employed for dichotomous outcomes while the inverse variance weighting method was used for continuous outcomes. I² statistic values were used to assess for heterogeneity with the following predefined thresholds: low (25%–49%), moderate (50%– 74%) and high (≥75%).47 Treatment effects were expressed as mean difference (with 95% CIs) for continuous outcomes and risk ratio (RR) (with 95% CI) for dichotomous outcomes. Primary analyses included data from both RCTs and observational studies, to greatly expand the available body of evidence and include a wider variety of ARF causes. Secondary analyses examined data from RCT studies only to increase the quality of evidence. Subgroup analyses for different types of respiratory illnesses (COVID-19, acute-on-chronic respiratory illnesses, hypercapnic ARF, chronic obstructive pulmonary disease, COPD and asthma), and sensitivity analyses of studies with low-moderate risk of bias were performed. All meta-analyses were performed by using Review Manager (RevMan, V.5.4; the Cochrane Collaboration).

Study characteristics and quality

A summary of included study characteristics is shown in online supplemental table 2, with more detailed information available in online supplemental table 3. COVID-19 was the most common cause of ARF (n=3782), followed by pneumonia (n=1583) (figure 1). Definitions of respiratory illnesses, as reported by the included studies, are reflected in online supplemental table 4. Overall, most studies included patients with mild-moderate degrees of hypoxaemia. Of the 17 studies that reported PaO2/FiO2 ratio as an inclusion criteria, 13 studies (76%) included patients with PaO2/FiO2≤300 mm Hg and 4 studies (24%) accepted patients with PaO2/FiO2≤200 mm Hg.

HFNO was most commonly administered via Fisher & Paykel Optiflow system (38 studies, 60%) and initiated at a median initial flow rate of 40 L/min (range 4–60 L/min, 34 studies). COT was initiated at a median initial flow rate of 10 L/min (range 2–15 L/min, 13 studies) and was most commonly delivered via nasal cannula (22 studies, 35%), followed by non-rebreather mask (20 studies, 32%) (table 1).

Risk of bias

Of 26 articles reporting RCTs (25 RCTs and 1 post hoc analysis), 20 (77%) had a moderate risk of bias and 2 (8%) at high risk (online supplemental table 5 and online supplemental figure 1). Five of nine randomised cross-over studies (56%) had a moderate risk of bias, with three (33%) at high risk (online supplemental table 6 and online supplemental figure 2). All five non-randomised interventional studies were of moderate-to-high risk of bias (online supplemental table 7 and online supplemental figure 3). For the 24 included observational studies (including 1 post hoc analysis), the overall risk of bias was low to moderate (online supplemental table 8 and online supplemental figure 4).

Outcomes

Table 2 provides a summary of all key outcomes for hospitalised patients, including the primary analyses of all study designs and secondary analyses of RCTs only. A more detailed summary can be found in online supplemental table 9.

Mortality

Mortality outcomes were reported in 17 studies (8 RCTs), with 5887 participants, for in-hospital mortality8 30 32 33 42 49 51–53 56–58 71 73 82–84; 20 studies (10 RCTs), with 6022 participants, for short-term (≤30 days) mortality27–32 34–38 40 42 48 50 51 58 59 61 72 and 10 studies (6 RCTs), with 5209 participants for long-term (>30 days) mortality.8 27 28 30–32 51 52 54 84 For short-term (≤30 days) mortality, 17 studies (85%) reported mortality at 28 days27–32 34–38 48 51 58 59 61 72 and 3 studies (15%) at 30 days.40 42 50 For long-term (>30 days) mortality, 7 studies (70%)8 30 32 51 52 54 84 reported mortality at 90 days and 3 studies (30%)27 28 31 at 60 days.

There was no significant difference between in the primary outcome of hospital mortality for patients receiving HFNO compared with COT (RR 1.08, 95% CI 0.93 to 1.26; p=0.29; 17 studies, n=5887, I²=37%) (figure 2). Secondary analysis of data from RCTs only showed similar lack of difference between the two groups for in-hospital mortality (RR 0.97, 95% CI 0.85 to 1.10; p=0.63; 8 studies, n=3031 participants; I²=0%)8 32 42 49 51 56 83 84 (online supplemental figure 5).

Request permission

Similarly, there were no significant differences in short-term mortality at ≤30 days (RR 0.94, 95% CI 0.81 to 1.09; p=0.42; 20 studies, n=6022, I²=53%) or long-term mortality at >30 days (RR 0.96, 95% CI 0.83 to 1.10; p=0.55; 10 studies, n=5209, I²=39%) between patients receiving HFNO compared with COT (online supplemental figure 6–7). Secondary analysis of pooled data from RCT only revealed a similar lack of difference between HFNO and COT for short-term (RR 1.00, 95% CI 0.86 to 1.17; p=1.00; 10 studies, n=3371 participants; I²=16%)31 32 35–37 42 48 50 51 59 and long-term (RR 0.96, 95% CI 0.84 to 1.09; p=0.50; 6 studies, n=2676 participants; I²=0%)8 31 32 51 54 84 mortality outcomes.

IMV escalation

Need for escalation from HFNO or COT to IMV was reported at variable illness time points (ranging from within 2 hours to the end of hospital admission) in 39 studies (23 RCTs) with 8932 participants6 8 27–37 42 48–62 67 71–73 76 77 81–84 (figure 3). Therefore, the overall need for IMV was defined as the need for IMV escalation at any time point during the study period. Most studies (n=21, 54%) reported IMV escalation at the end of hospitalisation or did not specify the time point(s) where IMV escalation was assessed. Eight studies (21%) reported IMV escalation at 28 days,27 31 32 34–36 51 54 four studies (10%) at 24 hours,8 67 72 81 four studies (10%) at 72 hours,50 56 57 72 one study (3%) at 30 days,42 one study (3%) at 48 hours6 and one study (3%) at 2 hours.60

Request permission

Pooled data from all study designs demonstrated a significant reduction in overall need for IMV escalation for HFNO when compared with COT (RR 0.85, 95% CI 0.76 to 0.95 p=0.003; 39 studies, n=8932, I²=56%) (figure 3 and online supplemental figure 8). For studies that reported specific assessment time points, a significant difference was observed (in favour of HFNO compared with COT) for IMV escalation at 28 days (RR 0.80, 95% CI 0.74 to 0.87 p<0.0001; 8 studies, n=2857, I²=0%) (online supplemental figure 9). No significant differences were seen for IMV escalation at 24 hours (RR 0.58, 95% CI 0.31 to 1.09 p=0.09; 4 studies, n=487, I²=0%) and at 72 hours (RR 0.66, 95% CI 0.42 to 1.03 p=0.07; 4 studies, n=276, I²=0%); however, these sample sizes were small (online supplemental figure 10–11).

Secondary analysis of data from RCTs only demonstrated similar results (in favour of HFNO compared with COT) in reducing the overall need for IMV escalation (RR 0.85, 95% CI 0.76 to 0.95 p=0.004; 23 studies, n=5046, I²=19%).6 8 31 32 35–37 42 48 49 51 52 54 56 59 60 62 67 76 77 81 83 84

NIV escalation

Need for escalation to NIV was reported in 16 studies (15 RCTs).6 8 31 32 35 37 48 54 56 60–62 67 77 83 84 Timing of NIV escalation was either reported at 24 hours (2 studies, 13%),8 67 at 28 days (3 studies, 19%)31 35 54 or unspecified (11 studies, 69%). Only one study (6%) defined NIV escalation as either non-invasive positive pressure ventilation (NIPPV) (bilevel support) or continuous positive airway pressure (CPAP).8 Four studies (25%) specified escalation to NIPPV only.31 32 54 84 The remaining 11 studies (69%) did not elaborate on NIV modality.

There was a statistically significant reduction in overall need for NIV escalation for HFNO compared with COT (RR 0.70, 95% CI 0.50 to 0.98; p=0.04; 16 studies, n=3076, I²=56) (online supplemental figure 12). Secondary analysis of pooled RCT data only yielded similar results favouring HFNO (RR 0.69, 95% CI 0.49 to 0.99; p=0.04; 15 studies, n=2974, I²=59).6 8 31 32 35 37 48 54 56 60 62 67 77 83 84

Inpatient admission rates

There was a small but significant reduction in the risk of needing inpatient admission for patients receiving HFNO in the ED compared with COT (RR 0.92, 95% CI 0.85 to 0.99; p=0.03; 3 studies, n=198; I²=0%)55 62 81 (online supplemental figure 13). This is likely due to the dominance (97.3% weightage) of an observational study55 which compared HFNO to COT in patients with PaO2<60 mm Hg in the ED setting. This small difference was not observed in the secondary analysis when data from RCTs only were pooled together instead (RR 0.88, 95% CI 0.53 to 1.45; p=0.61; 2 studies, n=77; I²=4%).62 81

Other hospitalisation-related outcome

Meta-analyses of all studies for all other outcomes, including ICU admission rates, ICU mortality, ED LOS, ICU LOS and hospital LOS, were not significantly different between patients who received either HFNO or COT (online supplemental figures 14–18). The statistical significance of these outcomes remained unchanged for secondary analysis of combined data from RCTs only.

Physiological and disability outcomes

Meta-analysis was not performed for change in PaCO₂ on arterial blood gases or disability scores (including dyspnoea, comfort and dryness scores) due to high methodological heterogeneity between studies. Of 31 studies (n=2019 participants) reporting a change in PaCO₂ levels between HFNO and COT, only nine studies (29%, n=766 participants)6 49 68 72 74 77 79 80 83 reported a significantly lower change in PaCO₂ levels in favour of HFNO compared with COT (online supplemental table 10).

Dyspnoea scores were most commonly measured using the Borg scale and visual analogue scale. 14 (56%), with 1601 participants, of 25 (n=3005 participants) studies7 31 32 49 54 61–65 68 74 78 81 reported significantly lower dyspnoea scores in patients receiving HFNO compared with COT (online supplemental table 11). Comfort scores were mostly reported using Visual Analogue Scale. 11 (46%), with 1088 participants, of 24 studies (n=3667 participants) reported significantly higher comfort or lower discomfort levels in the HFNO group compared with COT.6 7 35 37 54 62 64 67 69 77 81 The remaining 13 (64%, n=2579) reported no significant differences between the two groups5 8 31 32 48 49 51 60 61 63 66 70 80 (online supplemental table 12).

Subgroup analyses: types of respiratory illness

Online supplemental table 13 provides an overview of subgroup analyses results for different types of respiratory illnesses. In patients with ARF due to COVID-19, there were no significant differences between HFNO and COT in the primary outcome of hospital mortality (RR 1.06; 95% CI 0.87 to 1.30; p=0.54; 4 studies, n=2842 participants; I²=38%),30 32 33 42 short-term mortality (RR 0.87, 95% CI 0.73 to 1.04; p=0.12; 13 studies, n=4462 participants; I²=55%),27–32 34–38 40 42 or long-term mortality (RR 0.90; 95% CI 0.68 to 1.19; p=0.45; 5 studies, n=2923 participants; I²=58%)27 28 30–32 (online supplemental figures 19 A–C). However, HFNO therapy significantly reduced the overall need for IMV escalation compared with COT (RR 0.81, 95% CI 0.73 to 0.89; p<0.001; 12 studies, n=4384 participants; I²=34%)27–37 42 (online supplemental figure 19 D). In the secondary analysis of pooled RCT data only, there were no significant differences noted for all mortality outcomes and a non-significant trend towards reduction in the need for overall IMV escalation in favour of HFNO (RR 0.84, 95% CI 0.71 to 1.00; p=0.05; 6 studies, n=2297 participants; I²=35%).31 32 35–37 42 There were no significant differences between HFNO and COT for all other outcomes (including need for NIV escalation and LOS outcomes) in both primary and secondary analyses for patients with COVID-19 ARF (online supplemental figure 19 E–I).

Patients with acute exacerbations of chronic respiratory illnesses were identified in 12 studies (6 RCTs, 3 cross-over studies and 3 observational studies).73–84 Pooled data from all study designs demonstrated a significant reduction in-hospital mortality (RR 0.58, 95% CI 0.37 to 0.91; p=0.02; 4 studies, n=485 participants, I²=0%)73 82–84 for participants receiving HFNO compared with COT (online supplemental figure 20 A). This difference in-hospital mortality was not observed when pooled data were restricted to RCTs only (RR 0.39, 95% CI 0.06 to 2.54; p=0.32; 2 studies, n=375 participants, I²=0%).83 84 A significant reduction was also seen for ICU LOS (mean difference=−1.58 days, 95% CI −2.28 days to −0.89 days; p<0.0001; 1 RCT and one observational study, n=85 participants, I²=0%)82 83 (online supplemental figure 20 B). There were no significant differences between HFNO and COT for all other outcomes including need for IMV escalation, the need for NIV escalation and hospital LOS (online supplemental figure 20 C–E).

There were no significant differences between HFNO and COT in any outcomes for meta-analyses using the small number of studies limited to patients with acute hypercapnic respiratory failure, or patients with exacerbations of COPD or asthma (online supplemental figures 21–23).

Sensitivity analyses: studies of low risks of bias

Sensitivity analyses were performed excluding studies with an overall high risk of bias, defined as RCTs and randomised cross-over studies with a ‘high’ risk of bias using the ROB2 framework, non-randomised interventional studies with a serious risk of bias using the ROBINS-I framework or observational studies with scores <6 using the Newcastle-Ottawa Scale.

Using only pooled data from studies not at high risk of bias yielded the same findings for the primary outcome of hospitality mortality (RR 1.09, 95% CI 0.93 to 1.26; p=0.28; 16 studies, n=5817 participants; I²=40%),8 30 32 33 42 49 51–53 56–58 71 82–84 short-term mortality (RR 0.96, 95% CI 0.83 to 1.12; p=0.60; 18 studies, n=5915 participants; I²=52%)27–32 34–37 40 42 48 51 58 59 61 72 and long-term mortality (RR 0.96, 95% CI 0.83 to 1.10; p=0.55; 10 studies, n=5209 participants; I²=52%).8 27 28 30–32 51 52 54 84

Strengths, limitations and future research recommendations

This is the largest systematic review to explore the impacts of HFNO compared with COT on morbidity and mortality in patients with a broad range of respiratory illnesses. Limitations of this review include heterogeneous timing of outcome measurements, particularly the need for IMV escalation, as this varied depending on the time of treatment failure. Additionally, there was significant heterogeneity between studies in terms of administration of HFNO and COT, which were delivered in different countries and hospital clinical settings, using widely ranging flows and for varying amounts of time. Importantly, there was suboptimal reporting of information regarding the administration of HFNO (flow, time of usage, etc). Thus, it is not clear if the lack of change for many clinical outcomes was due to the intervention or suboptimal utilisation.

Second, the overall pooled population was small for certain outcomes, such as hospital admission rates. Third, the risk of bias was moderate for most studies and there was high heterogeneity for some outcomes such as hospital and ICU LOS. Thus, large, high-quality RCTs are required to further increase the power of meta-analyses to address several key clinical questions that remain unanswered. Fourth, it is challenging to blind participants and outcome assessors to the intervention (given the different sensation and delivery devices required for HFNO and COT), thus introducing bias, particularly for more subjective outcomes such as dyspnoea. Importantly, reporting regarding the causes of ARF in individual studies was suboptimal, with 3035 of 10 230 participants (30%) having no specified acute respiratory illness reported. Therefore, we were unable to better characterise the results by specific causes of ARF, which is crucial for understanding which patient populations with ARF are most likely to benefit from HFNO. It must also be highlighted that some of the meta-analyses were dominated by a few studies, including those performed for hospital and ICU admission rates, and the weaknesses of these dominant studies should be considered when interpreting the overall pooled results. Lastly, data from patients with a broad spectrum of respiratory conditions were pooled together to reflect real-world hospital settings for generalisability. While this introduces heterogeneity within the study population, it is important to examine HFNO’s clinical benefits compared with COT for patients with ARF secondary to heterogeneous causes as HFNO becomes more universally adopted.

Future studies that aim to determine which patient populations are most likely to respond to HFNO, measure outcomes at more standardised times, provide HFNO for longer durations and consider cost-effectiveness compared with COT are required.